Enable bf16 export for XNNPACK backend

[JakeStevens]

Three changes:

1. custom_sdpa stops force-upcasting to fp32.
2. A narrow fallback keeps older runtimes working.
3. bf16 XNNPACK delegation is requested when the model is bf16 as XnnpackPartitioner(enable_bf16=...) is opt-in

Note: enable_bf16 first appears in executorch==1.4.0.dev20260801

This means we cannot test with a stable version until a future branch cut post 1.4 for ET

Test Plan

1. Environment

conda create -n oe-bf16 python=3.11 -y
conda activate oe-bf16

2. Install

python install_dev.py --dependency_stack nightly

3. Verify the env

python -c "import executorch.version as v, torch, torchao, transformers; \
print(v.__version__, torch.__version__, torchao.__version__, transformers.__version__)"
# expect: 1.4.0.dev20260802+cpu 2.14.0.dev20260801+cpu 0.19.0.dev20260802+cpu 5.0.0rc0

4. Export

optimum-cli export executorch \
    --model google/gemma-3-1b-it \
    --task text-generation \
    --recipe xnnpack \
    --use_custom_sdpa \
    --use_custom_kv_cache \
    --qlinear 8da4w \
    --qembedding 8w \
    --output_dir hf_gemma3_bf16 \
    --dtype bfloat16

5. Check it runs

from transformers import AutoConfig, AutoTokenizer
from optimum.executorch import ExecuTorchModelForCausalLM

model_id = "google/gemma-3-1b-it"
model = ExecuTorchModelForCausalLM.from_pretrained(
    "hf_gemma3_bf16", config=AutoConfig.from_pretrained(model_id))
print(model.text_generation(
    tokenizer=AutoTokenizer.from_pretrained(model_id),
    prompt="Write a poem about machine learning.", max_seq_len=64))

Coherent output at ~11.9 tok/s: "The data flows, a river deep, / With patterns hidden, secrets sleep..."

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.